趋势解读：Perplexity's "Search as Code" lets AI models write，评估 LLM Agent 表现

Instead of calling a ready-made search API, models in Perplexity's new "Search as Code" architecture write their own search workflows as Python code. The company promises more precise results and lower token usage. Anyone who's watched an AI agent tackle a complex research task has seen the same pattern. The model writes a query, a search API returns a list of results, the model reads them, and then writes the next query. This loop repeats, often many times in a row. Perplexity calls this a bottleneck in a new technical report. Today's search engines were built for humans who want a neat list of blue links, but for an AI agent trying to run hundreds of searches in a few minutes, that setup is too rigid. The agent can only tweak the search term; everything else is a black box. "Search as Code" (SaC) changes that dynamic. Instead of calling the API, the model writes a custom Python script to run the search. The script runs in a secure sandbox, pulling from Perplexity's search backend. Basic operations like retrieving, filtering, deduplicating, and reranking are packaged as simple SDK functions. The architecture breaks down into three layers. At the top sits the model, which understands the task and decides on a search strategy. In the middle is the sandbox where the code runs. At the bottom is the "Agentic Search SDK," which breaks Perplexity's search engine into individual, mix-and-match functions. Standard search APIs are still there for quick questions. But for tough research, the model can go much deeper. It can fire off parallel queries, filter out the noise programmatically, and pull only relevant hits into its context window. According to Perplexity , that's where the win is. Standard search pipelines stuff an agent's context window with junk because the filtering logic is locked in. When the agent writes its own filters, the context stays lean, and the model keeps its bearings across long research sessions. To show how this works in the real world, Perplexity tested it on a messy cybersecurity task. An agent had to track down 200 critical software vulnerabilities (CVEs) published between 2023 and 2025. For each one, it needed to find the official vendor advisory, the affected software, and the exact version that patched the bug. News articles or blog posts didn't count. With SaC, the model wrote a three-stage script. It ran parallel searches tailored to how specific vendors like Mozilla or Google format their security bulletins. Next, it scanned its own findings, spotted the gaps, and ran targeted follow-up queries. Finally, it used a schema to verify that the CVE, product, and fix version all lined up. It worked. Perplexity says the agent nailed the task while using 85 percent fewer tokens than its standard pipeline. Competing systems got less than a quarter of the data right. Perplexity claims SaC beat rivals like OpenAI's Responses API and Anthropic's Managed Agents on four out of five benchmarks. The biggest gap was on "WANDR," Perplexity's own benchmark for broad research tasks, which it plans to release soon. Of course, take self-reported benchmarks with a grain of salt, but the comparison against Perplexity's own older architecture shows a clear, massive leap in performance. Perplexity frames SaC as part of a bigger trend. Traditional software relies on deterministic instructions. Frontier models add reasoning in token space. The most capable systems combine both: models for strategy, deterministic runtimes for batching and filtering, and search infrastructure as an I/O layer.